Hacker News Reader

Comment by roenxi | Hacker News Reader

I'm going to read the paper carefully and think hard, but going in to it I still have philosophical issues saying imperfect information is 'fundamentally' different.

Sure, the specific learning algorithm may not generalise from board games to poker. But to claim that Go is perfect-information is more theoretical than practical.

In practice, the 50% of moves your opponent makes are unknowns. Nobody can assess the entire game tree, so the fact that it is theoretically knowable isn't actually relevant in a game. Neither computers nor humans can compute it. Once the tree can be reasonably computed, between high level players the game is over (bar the occasional blunder).

View on HN

>> Nobody can assess the entire game tree, so the fact that it is theoretically knowable isn't actually relevant in a game.

That's not the point though. The point is that you only need to consider current state when considering your next move in Go or Chess. This is not the case in an incomplete information game, since the sequence of moves that led to the current state contains a lot of information. That's not the case in Chess or Go. Basically, Chess and Go satisfy the Markov Property, while Poker does not.

The current state in Go is the state of the board, the current state of poker includes some historical data. This difference is only significant to humans because we have quite poor memories.

An AI has perfect recall, and the fact that a variable is temporal really doesn't make a difference to the theory.

(Coincidental aside, this is an issue in Go as well due to the ko rule; but it isn't a major part of the game such as in poker.)

Which AI architecture has perfect recall, in a way that actually makes use of this information? Pretty sure AlphaGo etc does not make use of memory at all.

It's true that you can't model the entire game perfectly for any interesting game. But to play a game like Go you only need to know the current state in order to play optimally. If you play poker in a way that only accounts for your current state (i.e., you just take an expected value given the cards on the table, the cards in your hand, and the size of the pot) it won't take long for a competent player to learn how to consistently beat you. (Such a player would get tricked by bluffing, for instance.) In order to play correctly you need to model your opponent's strategy by understanding what their past moves were so that you can guess their probability of bluffing.

Reply on news.ycombinator.com